Stephen Fleming

Yes, that seems like a good thing to add!

It seems like `_read_legacy_10x_h5()` invokes ` _collect_datasets()` without taking `genome` into account? Perhaps this was lost during a refactor? https://github.com/scverse/scanpy/blob/bd06cc3d1e0bd990f6994e54414512fa0b25fea0/scanpy/readwrite.py#L222

Yes, I think line 222 should be ```python _collect_datasets(dsets, f[genome]) ``` I will make a PR, and I will try to add a couple of tests to highlight this issue.

@ivirshup or @flying-sheep or any other active maintainers, if you get a chance, could you consider the associated PR for this issue? It'd be great to have this fixed in...

@karenlawwc For the test.h5ad that you’ve saved using adata.write, I think you want to load it with sc.read_h5ad() rather than sc.read_10x_h5(), since the saved test file will be in AnnData...

@ivirshup @flying-sheep I wonder if this seems like the right fix to you?

Thanks for your comments. The PyTorch version I'm running is > pytorch 1.5.0 I typically see that we don't reach convergence until out past 100 epochs. Perhaps my early stopping...

As a note: I was able to successfully do this in python 3.10 ``` conda install -c conda-forge gfortran pip install ndd ``` and everything works fine. I was just...

For what it's worth, I do suspect the original issue was resolved when scvi was moved to the new pytorch-lightning backend. I haven't personally seen the issue since then, though...

What are we thinking about these changes? Look okay?